Research Intership project at UC Berkeley. For more details on the problem and our results, please check our report.
Objective: The objective was to build an intelligent traffic lights controller, capable of adapting the amount of green / red time for each lane as a function of the influc of vehicles.
Techniques: We model the traffic lights controller as a neural network trained via Reinforcement Learning methods. For this, a simple traffic simulator (Beats2) was used to simulate traffic and "rewards".
Conclusions: From our experiments, our adaptive traffic controller was observed to outperform fixed-time controllers by a large margin, thus providing a better solution to the problem of efficient traffic management.